Full Quiz · All Guides — The AI Learning Library

G1 · Question 1

What is the fundamental difference between an Agent and a Chatbot?

An agent uses a database; a chatbot doesn't An agent acts and loops with tools; a chatbot responds and stops An agent is always more accurate than a chatbot A chatbot can run code; an agent cannot

G1 · Question 2

What is the core flow of ReAct?

Plan → Execute → Review Read → Act → Terminate Think → Act → Observe (repeated) Prompt → Response → Evaluate

G1 · Question 3

Why does smolagents' CodeAgent write its actions as Python code rather than JSON?

Python is the only language the model supports JSON is too verbose for large payloads Python allows composability — loops, conditionals, and tool nesting in one step JSON requires an internet connection; Python doesn't

G1 · Question 4

What is "context compaction" in the context of a long-running agent?

Deleting old conversations permanently from storage Summarizing or dropping older turns when the window fills, to stay under the context limit Compressing binary data sent to the model API A technique to speed up model inference on long inputs

G1 · Question 5

What is the difference between memory and context in an agent?

They're the same thing — both refer to the conversation history Memory holds code; context holds text Context is the live window (finite, wiped between sessions); memory persists beyond it like a Session Note Memory is faster to access than context

G1 · Question 6

What problem does the Model Context Protocol (MCP) solve?

Making models write better code Speeding up inference on long prompts Standardizing how agents connect to external tools so capabilities are pluggable without bespoke glue per service Providing a graphical interface for running agents

G1 · Question 7

What should happen when a tool execution fails in a well-designed agent?

The agent crashes and must restart from the beginning The error is captured and fed back as an observation so the model can adapt (retry, change approach) The action is silently retried indefinitely A human must intervene immediately before any further steps

G1 · Question 8

Why is mini-swe-agent a good first agent to study?

It has the most features of any open-source agent It uses the most advanced frameworks available It strips away all scaffolding to show the bare loop — the model, not the framework, is doing the work It only works on simple tasks, making it easy to follow

G1 · Question 9

What is a Session Note in Mini-Agent?

A real-time log of all tool calls during the current session A persistent record of a run's learnings, saved to disk and reloaded in future sessions — a simple long-term memory A summary shown to the user when a session ends A configuration file for the agent loop settings

G1 · Question 10

Why does a coding agent need a sandbox?

To improve model reasoning performance To store code files permanently across sessions To isolate arbitrary code and command execution from the host system, protecting it from mistakes To provide a UI for the agent to interact with

G2 · Question 1

What is Claude Code's core agentic loop?

Prompt → Edit → Review → Merge Read → Write → Test Gather context → Take action → Verify results → repeat Plan → Code → Deploy

G2 · Question 2

Where should durable project rules be written so they survive context compaction?

At the start of every chat message In CLAUDE.md — read at the start of every session In a .env file in the project root Nowhere — Claude remembers everything automatically

G2 · Question 3

Which permission mode is read-only and best suited for reviewing changes to production code?

Default mode Auto-accept edits mode Plan mode Auto mode

G2 · Question 4

What does a Subagent do differently from the main agent in a Claude Code session?

It runs on a different model (always Haiku) It can edit files the main agent cannot It works in its own isolated context window on a scoped job, then returns a summary It runs multiple tasks in sequence before reporting back

G2 · Question 5

What is the key benefit of the Explore → Plan → Code → Commit workflow?

It makes Claude run faster by parallelizing tasks It reduces the number of files Claude needs to read It surfaces misunderstandings before any code is written — cheaper to fix a plan than a pile of edits It prevents Claude from needing to run tests

G2 · Question 6

What is "compaction" in Claude Code's context management?

Minifying code to reduce file sizes Deleting old sessions permanently to free storage Auto-summarizing older context when the window fills so the agent stays sharp A compression algorithm applied to files before editing

G2 · Question 7

What does a Hook do in Claude Code?

It captures the model's reasoning before each response It lets the user interrupt Claude mid-task It runs a shell command automatically before or after Claude's actions It connects Claude to external APIs via MCP

G2 · Question 8

What does MCP (Model Context Protocol) do in the context of Claude Code?

It controls how much context the model is allowed to use It manages the model's internal memory weights It connects external tools and data (Jira, Drive, Slack) to Claude via a standard protocol It is a protocol for compressing context before sending to the model

G3 · Question 1

What are the seven things a harness owns?

Models, weights, APIs, latency, throughput, cost, and reliability Task scheduling, context control, tool safety/environment, agent loop/collaboration, verification, self-evolution, and long-horizon execution Prompts, tools, memory, output, errors, logs, and tests Read, write, execute, connect, verify, compress, and terminate

G3 · Question 2

What is "progressive disclosure" in the context of the Skill System?

Gradually showing users more features as they learn the product Loading all skill documentation upfront so the model is always prepared Only loading a skill's name and one-line description at startup (~60 tokens); the full body loads only when the skill is actually triggered Revealing tool outputs to the user progressively rather than all at once

G3 · Question 3

What is the difference between Prefill and Decode in model inference?

Prefill writes the answer; Decode reads it back to the user Prefill is for short prompts; Decode is for long ones Prefill reads the entire prompt in parallel (compute-bound); Decode writes the answer one token at a time (memory-bandwidth-bound) Prefill caches the response; Decode sends it over the network

G3 · Question 4

What are the four failure modes of long-horizon tasks?

Timeout, memory leak, token overflow, and hallucination Model degradation, tool failure, cost overrun, and user abandonment Goal drift, context explosion, insufficient verification, and error accumulation Loop termination, cache miss, skill unload, and prefill cost

G3 · Question 5

What is "dreaming" in the context of agent memory management?

The agent hallucinating facts that don't exist in its knowledge base A metaphor for the model's creative reasoning capabilities Scheduled offline batch consolidation of memories (e.g. nightly) instead of risky real-time writes The idle state of a paused sandbox waiting for the next task

G3 · Question 6

What is a MicroVM (like Firecracker), and why is it used for agent sandboxes?

A very small Python virtual environment used to isolate dependencies A microservice that manages VM provisioning in the cloud A lightweight virtual machine with full VM-level isolation but near-container startup speed — safe and fast A compressed container image that reduces startup time by 10×

G3 · Question 7

Why do prefix-stable prompts matter for inference cost and speed?

Stable prompts are shorter, so they use fewer tokens They prevent the model from hallucinating An unchanged prompt prefix lets the system reuse cached attention data (KV cache hit), avoiding expensive recomputation They allow the model to batch multiple requests together

G3 · Question 8

What is the key difference between how Claude Code/Codex and Kimi Swarm handle multi-agent orchestration?

Claude Code uses Python; Kimi Swarm uses Go Claude Code runs locally; Kimi Swarm runs in the cloud Claude Code/Codex use predefined human-designed roles; Kimi Swarm has the model itself decide how to decompose and delegate work Claude Code supports more concurrent agents than Kimi Swarm